lab-1-recommendations-with-amazon-personalize

Last Updated: November 2019

Lab 1 - Personalize Your Recommendations

Overview

Amazon Personalize is a machine learning service that makes it easy for developers to create individualized recommendations for customers using their applications.

Machine learning is being increasingly used to improve customer engagement by powering personalized product and content recommendations, tailored search results, and targeted marketing promotions. However, developing the machine-learning capabilities necessary to produce these sophisticated recommendation systems has been beyond the reach of most organizations today due to the complexity of developing machine learning functionality. Amazon Personalize allows developers with no prior machine learning experience to easily build sophisticated personalization capabilities into their applications, using machine learning technology perfected from years of use on Amazon.com.

This lab will walk you through the following:

• Training, Deploying and testing your first recommendation model using Amazon Personalize
  • Downloading and preparing training data, based on the Movie Lens 100k data set
  • Importing prepared data into Amazon Personalize
  • Building an ML model based upon the Hierarchical Recurrent Neural Network algorithm (HRNN)
  • Testing your model by deploying an Amazon Personalize campaign
• Deploy and configure a Video Recommendation application
• Adding your campaign to Video Recommendation application

Setup your Jupyter Notebook environment

Train and deploy your first recommendation model using Amazon Personalize

1. If you have finished lab 0, you should now see the web interface of JupyterLab with our git repo cloned into it. Please click into the folder called lab-1-recommendations-with-amazon-personalize, and then onto our Lab 1 notebook, personalize_sample_notebook.ipynb.

   

Working Through a Jupyter Notebook

1. A notebook consisted of a number of cells; in SageMaker these will typically either be Code or Markdown cells. Markdown is used to allow for documentation to be defined inline with the code, giving the author a rich set of markdown formatting options. The first cell in this notebook, which is called Get the Personalize boto3 Client, is Markdown, and if you select any cell then the whole cell is highlighted.

2. The first Markdown cell describes what the following Code cell is going to do – for the sake of this lab you do not have to understand the code that is being run in the Code cell, rather you should just appreciate what the notebook is doing and how you interact with a Jupyter notebook.

3. To the left of a Code module is a set of empty braces [ ]. By highlighting the cell and then selecting the Run command in the menu bar, the Jupyter notebook will execute this code, outputting and code outputs to the notebook screen and keeping any results data internally for re-use in future steps. Do this now to execute the first code cell.

Note: if a Markdown cell is highlighted, then clicking Run will move the highlight to the next cell

3. Whilst the code is executing the braces will change to be [*], indicating that it is executing, and once complete will change to [1]. Future cells will have increasing numbers inside the braces, and this helps you see the order in which cells have been exected within the notebook. Directly below the code, but still within the Code cell, is the output from the code execution - this will include any error messages that your code has thrown. In this example, the code execurion successfully created the specified bucket in S3.

5. Now please continue to work through the notebook lab - read the comments prior to each Code cell in order to get an understanding as to what is going on, as these explain why we are doing each step and how it ties in to using the Amazon Personalize service.

Creating Parallel Solutions

Create Item-to-Item Similarities Solution

1. Using the same methods as before, go to the Services drop-down in the console and navigate to the Amazon Personalize service in another tab, and select Dataset groups. You will see the dataset group that you created earlier, and click on the name of your dataset group.

   

2. The left-hand side, which will show you the solution that you're currently creating via your notebook. Then, select Solutions and recipes, then click on the Create solution button.

   

3. Enter a suitable name for this solution, such as similar-items-solutions, select Manual recipe selection, then choose the aws-sims recipe and click Next - we don't need to change anything in the advanced configuration section

   

4. In the following screen just hit the Finish button and a new solution version will start to be created.

Create Personal Ranking Solution

1. Let's do exactly the same thing again, but this time we'll create a ranking solition. From the Solutions and Recipes screen that you are on, click Create solution, give it a name like rankings-solution, ensure it's a Manual recipe selection but this time select the aws-personalized-ranking recipe. Click Next and Finished as before

   

2. You now have three solutions being built off of the same dataset, and all three will slot into the application later.

3. You can jump back into the notebook at this stage, and skip down to the Create and Wait for Campaign cell, where you will see that we have now deployed a Campaign endpoint, have got some metrics from it, and made some recommendations.

Deploy the Video Recommendation App

Deploy the "Video Recommendation" Application

1. Whilst this application could be deployed anywhere, it uses both an EC2 Amazon Machine Image (AMI) and RDS Snapshot that have been stored in the North Virgina Region of AWS (us-east-1). Hence, please make sure that the Region selected in the AWS Console is alway US East (N.Virginia), as shown in the following diagram. The workshop will only function correctly if the EC2 configuration, CloudFormation template execution and SageMaker notebook are all using this AWS Region. AWS CloudFormation provides a common language for you to describe and provision all the infrastructure resources in your cloud environment.

2. The appication will run on an EC2 instance, but at some point we will need to connect to the server in order to carry out some configuration task. To do this we need to have an EC2 Key Pair configured on the server that you also have access to on your computer; hence, we need to create and download a new one. Click on EC2 from the list of all services by entering EC2 into the Find services box. This will bring you to the Amazon EC2 console home page.

3. On the left-hand menu scroll down until you see Key Pairs and select it, and in the resulting dialog click on the Create Key Pair button. This will bring up a Create Key Pair dialog, where you need to enter the name of a new key pair - call it myLabKey and hit Create. This should automatically download the file, or you may need to manually do so.

4. Optional - should you wish to later SSH in to your instance, you need to have your downloaded key-pair from earlier in an accessible location. It also must not be publicly readable, so if you are on a Mac or Linux system you can fix this with the following command run from the folder where you stored the key, remembering to replace myLabKey.pem with your key name!

$ chmod 400 myLabKey.pem

5. Click on the Services dropdown and select CloudFormation from the list of all services by entering CloudFormation into the Find services box. This will bring you to the Amazon CloudFormation console home page.

6. We are going to deploy a pre-built application via a CloudFormation template - this will be a fully-functioning recommendation system, allowing access to multiple Amazon Personalize features. But it has one drawback - there are no models built into it! So we will create them in this lab, and when they are ready we will re-configure this application to use them. But first we need to deploy this skeleton application but downloading this file from the workshop repository. Right-click on the following link and download the template to a file on your local computer, remembering to keep it as a text file with a .yml extention.

https://raw.githubusercontent.com/johanneslanger/recommendations-on-aws-workshop/master/assets/cloudformation_template.yml

7. Click on the Create Stack button to start the deployment wizard, and in the Specify template section select Upload a template file, click on the Choose file button, and select the template file that you just downloaded. Then click on Next.

8. The next screen asks for more configuration parameters, but only a few of these are required. Enter you values for the following, then click Next when you're done - you are free to enter new values for the DjangoAdminLogin or DjangoEmailAddress, but the defaults for these and other parameters are fine.

9. Stack name - enter something simple, as as LabStack

10. Parameter - DjangoAdminPassword - enter a password of your choice

11. Parameter - KeyName - select your previously-defined EC2 kay-pair

12. There then follows two more screens. The first is called Options, but we have none to enter so just click on Next. The second is the final Review screen - verify the parameters, hit the checkbox next to the statement "I acknowledge that AWS CloudFormation might create IAM resources with custom names." and then click on Create stack. This will then go and create the environment, which will take around 10-15 minutes minutes. Unfortunately, we are creating IAM resources, so we cannot continue until it has completed - so read ahead and get a feel for what's coming up next.

Using the Video Recommendation App

Notes on the Application

1. The RDS database is postgres, and we have included the pgcli tool with this deployment, as you may wish to look at the database schema structure, examine the data that was in the RDS snapshot, or potentially update this all after you have customised the Django application for your own needs. There is a startup script for this in the /home/ec2-user/personalize-video-recs/videorecs/ folder, which is the root of the whole Django framework application.
2. CURRENT BUG - when the CloudFormation template is complete there is a running web-application, but for some (currently unknown) reason it fails its health check. However, SSH'ing into the instance will show the web-server job running, configuration is correct, etc. If you get a Bad Gateway error when you navigate to the application, please terminate the EC2 instance - auto-scaling will bring another back to life, which always seems to be fine.

Running the Video Recommendation App

1. The application server should already running running. You can find the two URLs required for different aspects of the app in the CloudFormation outputs panel.

2. The URL of the application is your ALB followed by the /recommend/ path, although there is also an /admin/ path configured that we'll use later. For now connect to your application using the AppEntrypoint URL

3. You should see the following screen in your browser - no Model Precision Metrics are available, as we haven't added any models yet to the application. You can also see that documentation for this is present, but be aware that it may not be 100% up to date with coding changes on the demo.

   

4. If you hit Select Random User then you'll be taken to the main Recommendation screen, which starts by showing you a random user's top-25 movie review titles. However, you'll see on the Model dropdown on the left that there are no models available, and if you change the Personalize Mode to either Personal Ranking or Similar Items then it's the same story - you can see the movie reviews, and most-popular titles in a genre, but no recommendations. We need to get the solutions and campaigns built in the notebook, then you can come back and plug in the models.

   

At this point we require the solution that is being built in the notebook to complete and the associated campaign to have been created - until that time we cannot move forward, so you may wish to get some refreshments if you are still waiting for those two steps to complete.

Create Additional Personalize Campaigns

If you have built the additional two Personalize models, for Item-to-Item Similarities and Personal Rankings, then you'll need to create the associated campaigns for these solutions, as it is the campaigns that we will add to the application. If those solutions have been built then continue with these steps, but if not you can always come back to these steps later before adding them to the application.

1. In the AWS Console, go to the Amazon Personalize service console, click on Dataset groups link on the left-hand menu, and select the personalize-recs-dataset-group link, then click into the Campaigns menu item on the left. Select Create campaign

   

2. First, we want to build the campaign for the Similar Items model - enter a name for the campaign, such as similar-items-campaign, select via the drop-down that solution that you previously built in the console, similar-items-solution, and ensure that minimum TPS is set to 1. Hit Create campaign

   

3. Now build the campaign for the Personal Rankings model - follow the same steps as before, but this time use rankings-campaign for the campaign name and select the rankings-solution model in the drop-down control.

Plug In the Recommendation Model(s)

The application uses the Django Administration feature to define models that are available to the application. This allows multiple models of different types to be configured, and injected or removed from the application at any time. There are three modes of operation of the application:

• Recommendations - standard recommendations, allowing different 2 models to be compared at once
• Personal Ranking - re-ranks popular films in a genre, with a single model on-screen at once
• Similar Items - shows items similar to others, with a single model on-screen at once. You can optionally send this list through a Personal Ranking model if you have one defined

Each of these modes allows multiple models of their type to be used, but each mode can only show one or two different models simultaneously - however, you can choose any configured model at any time.

1. Login to the Django Administration site. This is at the same URL as the main application, but replace /recommend with /admin at the end of the URL, as shown previously in the CloudFormation Outputs panel. This will bring up the following screen, so login now with the credentials that you supplied when you ran the CloudFormation template:

   

2. This brings up the Site Administration screen, which show entries for Groups and Users (which we don't need), but also a section called Recommend where you can add Personalize models to the app. Click on +Add link to begin to add a new model

3. Back on the AWS Console, go to the Amazon Personalize service console, select the personalize-recs-dataset-group and then on the left-hand menu click Campaigns. This will show your personalize-lab-recs-campaign, along with the campaigns for the other two solutions if you've created them. If you've created all three then you should see something like this, but for your other two campaigns may already have been created

   

4. Click on the personalize-lab-recs-campaign and you'll see the Campaign ARN - copy this, and head back to the admin screen. Enter Personal Recommendations for the model name, enter the ARN where it asks, ensure that the Model type is set for recommendations and set the Model sort order to 1. Click on SAVE to save the definition.

   

5. The application will use the sort order field to decide how to order models in the on-screen drop-downs. Only models of the right type are shown on the relevant screen, but there is no validation that you have entered the correct model type, and if you put a SIMS model on the Rankings screen then the application will throw errors.

6. If you also have a SIMS or Personal Ranking campaign then go ahead and add them now in the same way - if they haven't yet completed then you can come back and add them later. You can then close the admin screen and head back to the main application web page

7. The main screen now shows the three models (or maybe just one) that we've built - it lists the precision metrics for each one, and as you add or remove models from the Django Administration page the changes will be reflected here. Now click on the Select Random User button

   

8. The screen will look as before, but now if you click on the Model 1 drop-down you will see that our one Recommendation model is present - if you select it then the screen will refresh to show recommendations for this user using that model.

   

9. You can step through users to see how these look for different demographics of users. If you had mutiple Recommendation models defined then they would also be in the two model drop-downs, and you'd be able to show two completely different recommendation models, based upon different user demographic or item metadata, allowing you to compare and contrast different approaches.

10. Try out the Personal Ranking personalize mode - this takes a list of the most popular movies in the dataset, either as a whole or in just a single genre. This will take that list and re-rank it into an order for this particular user, ensuring that the ones that are most likely to be interested in are shown first.

11. Finally, try the Similar Items personalize mode - this starts with the user's top-ranked film, and finds a list of films that people who watched this also watched. This is done without reference to the user's preferences, and the list is generated based upon what's in the dataset as a whole. However, if you also have a Personal Ranking model defined then the Ordering drop-down will re-rank this list into one that is more suited to the user's preferences.

Additional Campaigns to Build

If you look at the embedded documentation you'll see that it talks about 3 other models, which there isn't time to build during this Lab. They involve the user of additional data files - a user demographic file, and a item metadata file, all of which are supplied with the Movie Lens data set in your Sagemaker Notebook. Because they required additional data-sets, you need to create each of these within their own Personalize Dataset Group, and you also need to re-import the original interactions file DEMO-movie-lens-100k.csv that you uploaded into S3 during the notebook - this is because Personalize trains solutions on all data files witin the Dataset Group.

The three models that you should build are as follows:

• Using a USERS file, create a model that takes into account user's demographic details such as age, gender and occupation
• Using an ITEMS metadata file, create a model that also takes into account the movie year and the top-4 genres associated with that movie as 4 separate metadata fields
• Using an ITEMS metadata file, create a model that also takes into account the movie year and then compounds the top-4 genres into a single metadata field

Observations are that demographics are absolutely not a good indicator for movies recommendations, nor for things like book recommendations - this isn't an issue with Amazon Personalize, rather it is a know issue with using age and gender to predict likes and dislikes of media. Also, the single, compound genre certainly seems more accurate for the first 5 or 10 responses, but for the set of 25 response as a whole the multiple genre model probably gets a better list of movies than the compound one.

Closing Down Your Resources

Terminating the Notebook Instance

1. Open the Amazon SageMaker console and click on Notebook instances
2. Find the notebook instance listed as [Name]-lab-notebook, select its radio button and then click the Actions dropdown.

3. Click Stop to stop the Notebook Instance. This does not delete the underlying data and resources. After a few minutes the instance status will change to Stopped, and you can now click on the Actions dropdown again, but this time select Delete.

Note that by selecting the name of the Notebook instance on this dialog you are taken to a more detailed information page regarding that instance, which also has Stop and Delete buttons present – notebooks can also be deleted using this method.

Conclusion

Upon completion of this lab you will have performed the following:

• Launched a Jupyter notebook from with the Amazon SageMaker service
• Imported external files into the notebook environment
• Used the pandas libraries to do some pre-processing of the source data
• Built and deployed an ML model based upon the HRNN algorithm
• Tested your model via just a few lines of code
• Deployed your model into a live application

You should now be able to embed this model from within your own application code, using any language that is supported by the AWS SDK. Happy recommending!

Feel free to continue with Lab 2 - Training a custom recommender model using Amazon SageMaker